Debris sound effect
The sound of falling debris and small ruble is an interesting failure case for our model. Figures \ref{819959} and \ref{928895} show real and generated waveforms for a falling debris sound effect. Figure \ref{667351} shows a side by side comparison of the corresponding spectrograms. In the real waveform we see several clear and distinct audio events, which correspond to the hits of small individual pieces of rubble after an initial impact. However, in the generated audio things are a lot more muddied. In general, the generated audio waveform contains too much variation in the higher LODs, resulting in a much higher level of baseline noise. Individual hit events are not clearly distinguishable from this noise and don't have clearly defined structure. In the real waveform, it is clear that each hit event is very structurally similar to previous hit events, however, in the generated waveform we see little semblance of any repeated patterns. Instead we observe what seem more like random spikes in amplitude. We also note that in the real sample we can see from both the waveform and spectrogram that the sound gradually decreases in volume over time and we also begin to lose some frequencies. However, in the generated sample this volume decrease and loss of specific frequencies is much less apparent. The resulting generated sound effect subjectively appears to get the initial impact correct, but fails to capture the long running patterns of repeated small debris hits, and instead we get something that sounds more like random scratching and popping. It is clear from the spectrogram representations however, that our model is able to correctly capture the initial spectral characteristics of the sound, generating sound effects in the same spectral range as the original. What appears to be missing is the time dependent structure and repetition.